Intelligent pest forecasting with meteorological data: An explainable deep learning approach

Abstract

Accurate pest outbreak forecasting is crucial for protecting agricultural yields and environmental health. This study addresses limitations in existing meteorological data-driven approaches by introducing a novel, interpretable deep learning architecture: the Attention-based Long Short-Term Memory Interaction Convolutional Neural Network (ALIC) model. The ALIC model excels at extracting intricate interrelationships within multivariate time series data encompassing both meteorological and historical pest data. The proposed architecture employs a meticulous three-stage processing pipeline. The initial stage, a feature attention layer, pinpoints the most influential forecasting factors within the input data. Subsequently, the feature interaction extraction layer leverages Long Short-Term Memory (LSTM) networks to capture temporal dynamics and a novel Interaction Convolutional Neural Network (ICNN) to extract feature interactions between meteorological and historical pest data. Finally, the output layer generates accurate pest outbreak forecasts. The ALIC model’s effectiveness is rigorously validated using real-world data obtained from observation stations in Huiyang and Shantou, Guangdong Province, China. The model demonstrably outperforms benchmarks, achieving significantly lower forecasting errors across diverse forecast horizons. This finding underscores the ALIC model’s efficacy in capturing intricate relationships between pest outbreaks and meteorological variables. Importantly, our feature importance analysis pinpoints temperature, rainfall as crucial for pest forecasting, bolstering the model’s utility in pest management.